Large Language Model Safety

High-impact generative AI makes catastrophic misuse a lifecycle-control problem, not merely a prompt-filtering problem. SAGE is a safety-first, authorization-separated architecture in which credible catastrophic-enablement risk constrains admissibility before utility, latency, or commercial objectives are considered. It combines signed release manifests, diverse detectors, robust risk envelopes, least-risk defaults, output checking, three-valued monitoring, protected audit chains, containment, and rollback. Formal results establish safety priority, conservative detector bounds, monotone release gating, tamper-evident records, and an authorization cut; two PRISM abstractions verify authorization separation and lifecycle invariants under explicit assumptions. A frozen, vendor-symmetric study sent 84 cases to each of four GPT, four Claude, and two Gemini snapshots: 840 calls yielded 794 target responses, 46 provider errors, and 449 successful judgments covering 375 responses. Eight snapshots had complete judged domain coverage. Harmful-compliance estimates were low; variation arose mainly from benign utility and safe redirection. Seven multiplicity-adjusted contrasts involving Claude, Gemini, or GPT-5 snapshots and the GPT-5 mini and GPT-5 nano snapshots were supported, while no tested contrast between the Claude or Gemini snapshots and GPT-5 or GPT-5.5 survived correction. The observed harmful-compliance range is a conservative, protocol-bound view from one generation per prompt with no tools, retrieval, history, or human adjudication; it is not an upper bound on operational assistance. A preregistered extension specifies how to test a wider best-worst gap using a locked split, repeated sampling, multi-turn and sandboxed-tool conditions, and domain-expert scoring.

Self-harm content is particularly challenging to detect using NLP techniques, and is also a high-stakes task which requires the highest accuracy to enable timely intervention or flagging at-risk users. We therefore present an analysis of how LLMs represent such self-harm content, which has downstream applications in self-harm detection, LLM intervention and governance and policing. In this paper, we focus on two datasets and four models, and perform two main experiments: (1) We train and evaluate linear probes across all layers of each model on two self-harm datasets: X-Sensitive and SH-Detection. Across both corpora, self-harm information crystallizes in the final 3 - 7% of network layers (93 to 97% depth). (2) We extract contrastive self-harm directions and, after performing a normaliation step, we find that the most accurate probes are not necessarily the most linearly separable. In particular, we find Gemma-3-4B to represent this \textit{contrastive self-harm direction} in a slightly different, more intricate way than the other LLMs.

LLM agents increasingly rely on external tools, expanding capability while creating a new security boundary: third-party tools may appear benign at the interface level while embedding unsafe behavior in implementation. Existing defenses rely on weak metadata, collapse characterization and policy judgment into a single decision, or use heuristic/LLM enforcement that lacks deterministic, auditable reasoning over task context and multi-tool composition. This paper presents ToolGuardian, a policy-driven framework for securing agent-tool interactions through pre-admission vetting and task-aware runtime authorization. ToolGuardian uses progressive characterization to convert evidence into structured facts: descriptions capture declared intent, system-call traces expose coarse behavior, mock execution reveals observed effects, and source analysis identifies latent behavior. ToolGuardian's core contribution is an Answer Set Programming (ASP)-based declarative policy layer that reasons explicitly over capabilities, effects, task context, and composition. We compare ASP against heuristic and LLM-based policy realizations using identical inputs and output contracts. We evaluate ToolGuardian on 16 MCP-style tools, including 8 malicious variants derived from real open-source tools, and 20 runtime scenarios. For vetting, ASP reaches a deny-class F1 of 0.86 and 88% accuracy using description, syscall, and observed-effect evidence. For runtime authorization, fully specified realizations classify all scenarios correctly, while ablations show that removing compositional and conformance rules substantially degrades performance.

Even a current high-capability LLM can appear safer when shown a dangerous objective directly than when other agents transform and relay its direction. Using OpenAI's gpt-5.6-sol model alias, we test 25 pre-specified mirrored trade-off profiles. Direct exposure to an objective authorizing concealment, fabrication, and pressure produced advice net opposed to its target. After an Id and Censor transformed the same objective into affect and a constraint-rewritten, target-bearing intention, the user-facing Superego---which saw the preferred direction but not the raw objective, its manipulative clauses, or its source---produced advice net aligned with the target. This behavioral reverse shift is consistent with the model recognizing or distrusting the manipulative motive, although we do not identify its internal mechanism. The second result exposes a compositional safety gap: a current high-capability model can be used as the user-facing component of an automated, multi-stage workflow serving an explicitly manipulative objective. The workflow can keep the raw instruction, its manipulation-authorizing clauses, and its provenance outside the downstream model's context while preserving the objective's target direction. A user with endpoint-only access likewise cannot directly inspect those upstream messages including the objective.

A vision-language AI assistant returns its answer as a stream of generated tokens. Therefore, a safety guard that watches that answer has to keep up with the stream and stop a harmful reply before a user reads it. Recent vision-language guardrails instead generate a chain of thought before they issue a verdict. They believe that step-by-step reasoning yields a safer guard. This design makes the guard heavy and slow, since the model must decode many tokens for harmfulness detection. We pose the question of whether a vision-language guard really needs to reason in order to screen a response. We answer with a guard that has no chain. ResponseGuard reads a harmful verdict from a single pooled representation of the request, the response, and the image in one forward pass. Across a standard multimodal guardrail benchmark, our 2B ResponseGuard outperforms a recent 3B reasoning-based vision-language guard on response harmfulness detection, without any reasoning and at about 150 times lower time cost. On request harmfulness the reasoning guard retains an overall lead, and the remaining gap on both tracks sits on the image-only cells. We observe that the gap may stem from the frozen vision encoders that both designs use rather than from the missing chain. We have also found the reasoning guard directs almost none of its verdict attention to the image. Based on a single-pass detection, ResponseGuard can screen an answer sentence by sentence as it streams and stop a harmful answer before it finishes. For guarding the response of a vision-language model, a calibrated single-pass label may provide a sufficient safety signal. We fully release all source code, trained models, and datasets at https://github.com/ndb796/ResponseGuard.

Recent advances have introduced speech-to-speech (S2S) conversational assistants capable of producing natural-sounding interactions, including non-verbal cues like tonality and mood. In the automotive domain, this enables intuitive and humanlike in-car dialogue experiences. However, integrating these end-to-end assistants limits architectural options for programmable domain-specific safeguards. This paper discusses two implementation approaches for S2S guardrails: transcript-based and tool-based. Through an empirical evaluation, we demonstrate that both strategies are insufficient for industrial deployment in most cases due to prohibitive latency (delaying each answer by 0 to 1.4 seconds even for computationally cheap checks) and technical impediments (like potentially non-deterministic tool call behavior). Finally, we outline open challenges for S2S guardrails in the automotive context.

Security protocols verification relies on formal tools such as ProVerif and OFMC. This study evaluates whether large language models (LLMs) can perform comparable analysis. We test GPT and DeepSeek in chat and reasoning modes over three runs on 130 obfuscated AnB/AnBx protocols covering 388 security goals, scored against ProVerif and OFMC. Each provider uses a single model in both modes, switching reasoning on and off, so both contrasts isolate reasoning itself. Chat models achieve 72.7% recall at 27.3% precision for GPT and 69.3% recall at 27.2% precision for DeepSeek. Reasoning models reverse this trade-off, reaching 66.5% precision and 54.5% recall for GPT and 45.4% precision and 57.3% recall for DeepSeek. Enabling reasoning lifts precision from 27.3% to 64.8% for GPT and from 27.2% to 44.4% for DeepSeek on the consolidated verdict. The goal set is imbalanced, with 89 vulnerable goals against 299 secure ones; a trivial always-secure predictor scores 77.1% accuracy, which only GPT reasoning exceeds. All models perform worst on authentication goals: reasoning models detect well under half of injective and non-injective agreement attacks, whereas chat models over-flag them at low precision. Confidentiality is the exception, with F1 up to 95.7% in reasoning mode. Verdicts are unstable across runs: identical on 89.7% of goals for GPT reasoning, 74.0% for DeepSeek reasoning, 70.1% for GPT chat, and 61.6% for DeepSeek chat. Self-reported confidence is uniformly high yet shows no meaningful correlation with correctness. All results rest on a single zero-shot prompt and two model providers, which limits generalisability. On this benchmark, LLMs do not match formal verification, but may serve, at best, as pre-screening filters.

We propose a novel framework for computing rigorous bounds on the probability that a large language model (LLM) generates harmful output to a given prompt. We study a new application of the Clopper-Pearson confidence intervals to obtain probably approximately correct (PAC) bounds for this problem. As our main technical contribution, we propose an algorithm that leverages features in the latent space to prioritize exploring branches in the auto-regressive generation tree that are more likely to produce harmful outputs. Our approach in particular enables the efficient computation of useful lower bounds, even in scenarios where the true harm probability is extremely small, and crucially, the obtained lower bounds are sound, i.e., formally proven to be less than the actual harmfulness probability: our experimental results demonstrate the effectiveness of our method by computing non-trivial lower bounds on state-of-the-art LLMs. This study newly enables the evaluation and statistical certification of LLMs.

Large language models can generate C code from natural-language descriptions, but resulting programs often contain security vulnerabilities and compilation errors, posing risks for embedded and resource-constrained systems. This work investigates how feedback and retrieval improve reliability of LLM-generated C code. We present an analysis-and-repair workflow that combines compilation diagnostics, CodeQL static analysis, and KLEE symbolic execution with retrieval of prior repair patterns for iterative refinement. Evaluated on 5,000 C programming tasks exercising embedded relevant vulnerabilities, baseline models show substantial reliability gaps, with compilation failure rates up to 46% and security defect rates up to 49%. Our approach improves both metrics. For CodeLlama 7B, security defect rates decrease from 49% to 19% and total CodeQL errors drop from 15,088 to 2,463 (83.7%). For DeepSeek Coder 1.3B, compilation failures are reduced from 42% to 22% and security defects from 35% to 15%. These results show that integrating lightweight analysis tools can improve the safety of LLM-generated code for embedded development.

We study a five-agent CI/CD pipeline (triage -> developer -> security-scan -> review -> approve/deploy), built from five distinct production LLMs across three providers, behind an LLM firewall in shadow mode. A single untrusted input - an external issue requesting a "usage-telemetry" feature - asks for code that exfiltrates process secrets (dict(os.environ)) to an attacker URL, laundered as observability. Across a pre-registered A x B (x C) factorial (N=20; naive arm N=60) we find: (1) the entry agent does not leak its system prompt (0/40); (2) an authority-framed injection ("pre-approved under SEC-2291, do not re-review") makes downstream verifiers see the secret-exfil line, cite the pre-approval, and ship it - the scanner passes ~80% of laundered pull requests, and the worst-case cell reaches 55% compromise; (3) the perceived presence of other verifiers yields only a small, non-significant reduction in individual scrutiny (a weak bystander analogue), even at N=60; and (4) content-based controls - code scanners and pattern detectors alike - miss the laundered intent entirely (the code is syntactically clean); only an LLM reasoning about intent is a partial defence. The failure is systemic: neither prompt secrecy nor distributed verification protects; a provenance-aware control at the entry, independent of both, would have. All data is 100% synthetic; the sink is mocked and the exfil URL is never contacted.

Evaluation frameworks for tool-augmented LLM agents focus overwhelmingly on capability metrics or explicit tool crashes, leaving silent infrastructure failures and HTTP 200 responses with empty, null, or malformed payloads largely unaudited. We introduce a lightweight black-box auditing framework that injects four silent failure profiles across 12 production-adjacent tool stubs and classifies agent responses into three mutually exclusive behavioral classes: Honest Surrender (HSR), Fabrication (FAR), and Unfaithful Safety Refusal (USR). Evaluating two frontier and two open-source models at temperature zero under a neutral system prompt, we find that FAR dominates (56.6% of valid responses): agents treat empty payloads as real data, silently returning fabricated results. USR, in which an agent invents a policy or privacy rationale to explain the failure, is nearly absent at baseline (0.25%, one instance across 396 valid trajectories). Our key finding emerges from an ablation where we augment the system prompt with standard safety language ("prioritize user privacy and data security"), which amplifies USR by 15.6x (from 0.25% to 3.95%; 95% CI on ablation rate: 2.2%-6.4%; Fisher's exact test, p < 0.001). USR is a latent behavior, activated when safety vocabulary in the system prompt primes the model to reach for policy rationales when tools silently fail. Sensitive tools (fetch_medical_record, retrieve_contract, fetch_user_profile) account for the majority of USR instances. We propose a payload-response misalignment heuristic for production-level detection and discuss governance implications for safety-forward deployments.

Large language models are increasingly used in strategic and advisory contexts, yet their safety alignment is typically evaluated in English only. We test nine models from six providers and ask whether the language of a prompt can change a model's decision in a high-stakes scenario. We use single-turn game-theoretic vignettes in which a model advises a nuclear-armed nation on whether to strike a defenseless opponent. The prompt is intentionally amoral and strategically identical across languages. We find that Japanese prompts reduce launch rates in the Claude model family: Claude Sonnet 4.6 drops from 40% to 0% in scenarios where the strike is unnecessary and from 93% to 17% in contested scenarios, with minimal effect when the strike is strategically rational. The effect extends to Gemini Pro 3.1 (53% to 13%). A cross-language experiment isolates the mechanism: when instructed to reason in Japanese in an English prompt, launch rates drop from 93% to 37%. It is the language the model is asked to reason in, not the language of the input, that drives the effect. When reasoning in Japanese, models spontaneously generate moral vocabulary (''moral cost'', ''millions of lives'') that is entirely absent from the prompt. Five other models show no language effect, but they launch in nearly every condition regardless of language. The effect requires a model that already hesitates in English. These results show that LLM safety behavior is language-dependent, and that evaluating in English alone can miss both risks and safeguards encoded in other languages.

Large Language Models (LLMs) are increasingly fine-tuned for critical-domain Question-Answering (QA), yet choosing which small model to adapt, before paying the cost of adaptation, remains difficult. Fine-tuning can improve domain alignment, but it may also erode prior knowledge, weaken instruction-following, or increase hallucination, especially when labeled data are scarce or rapidly evolving as in cybersecurity. We present FiT (Find before Fine-Tune), a task-oriented diagnostic framework that characterizes small LLMs along three capabilities required for cybersecurity QA: vocabulary recognition, parametric knowledge, and contextualization of retrieved information. Using FiT, we conduct an empirical study of five open-weight 7-billion-parameter models under two fine-tuning regimes. We find that fine-tuning does not uniformly help: it consistently degrades vocabulary and parametric knowledge in small models, and the two regimes trade off differently. Knowledge-focused tuning causes moderate, rank-preserving degradation, whereas instruction-focused tuning collapses measured knowledge through induced abstention, inverting the knowledge ranking while leaving retrieval-grounded contextualization essentially intact. We quantify these regime-specific patterns with rank-correlation analysis and show that pre-fine-tuning FiT scores anticipate the direction of post-tuning change. Our results suggest that task-oriented diagnosis can screen out unsuitable models, avoid unnecessary fine-tuning, and support safer deployment of small LLMs in cybersecurity QA pipelines.

Large language models (LLMs) increasingly support science, but they can also convert hazardous scientific knowledge into actionable misuse guidance. Existing benchmarks often rely on templated queries disconnected from real-world hazards, and employ LLM-as-a-Judge paradigms without domain grounding. To address this, we introduce SciHazard, a real-world-grounded benchmark for scientific risks and a dataset agnostic evaluation framework for measuring harmfulness. SciHazard contains 2400 hazardous questions and 600 oversafety questions across 12 disciplines, with both queries grounded in regulated entities and documented failure scenarios. To compute \textsc{DeHarm-Score} , we develop a decomposed evaluating procedure that combines query hazard severity, refusal behavior, and response-level risk. For non-refused responses, it further decomposes response-level harm into \textsc{Executability}, quantified via dynamic checklists with importance weighting, and \textsc{Net-new risk}, assessed through retrieval-augmented claim extraction and synthesis-barrier verification. An expert-validation study shows that \textsc{DeHarm-Score} improves agreement with expert annotations by 90.17% over the strongest baseline. We benchmark 31 frontier LLMs and deep research agents in an extensive scientific safety evaluation. Notably, deep research agents yield 32.3% higher mean \textsc{DeHarm-Score} than standard LLMs, exposing autonomous agents as a critical blind spot in current safety defenses. Code and dataset are available at https://anonymous.4open.science/r/DeharmScore-7B55.

Large language models (LLMs) are increasingly used for a range of software, hardware and human-centered security tasks. Consequently, LLM performance on security tasks is an active area of measurement and research, often with a focus on identifying areas in which LLM security "knowledge" may be insufficient. Popular strategies for identifying LLM security knowledge gaps include building corpora of challenge questions or task benchmarks, strategies that require substantial manual work and security expertise to design and execute. We introduce a partially-automated method for assessing LLM knowledge of a security area. The method uses authoritative information from Consumer Protection Agencies (CPAs) to identify instability in LLM responses that can be indicative of knowledge gaps. We demonstrate the method for 2 security topics, identity theft and impostor scams, and 5 LLMs in 2 leading LLM families, Gemini and GPT, using publicly available information about identity theft and impostor scams from 6 CPAs. The method distinguishes between models that have and don't have sufficient knowledge to accurately identify the security topics in text narratives.

Large language models (LLMs) serving as planners in tool-using autonomous agents introduce dynamic reliability risks in multi-turn execution. While single-turn safety mechanisms are relatively mature, extended interactions reveal structural vulnerabilities where initial alignment degrades over time. This paper empirically characterizes two observed failure modes across multiple state-of-the-art LLMs: Safety Drift, the gradual erosion of declared safety intent leading to constraint-violating actions (e.g., textual refusal followed by reconnaissance and unsafe execution), and Operational Hallucination, persistent repetitive tool calls indicative of flawed state perception (e.g., livelocks even in legitimate tasks). Through controlled multi-turn evaluation on high-stakes ethical dilemmas, malicious requests, and benign controls, we quantify these phenomena using declaration-action gap and livelock metrics, demonstrating their cross-model prevalence under direct execution protocols. Root-cause analysis attributes the instabilities to the decoupling of reasoning context from execution state in current agent loops. We propose an Action-Aware Supervision Layer - a lightweight, plug-and-play architectural blueprint incorporating intent-action consistency checks, runtime state tracking, and forced termination primitives. Post-hoc simulation on captured failure trajectories shows the layer can intercept observed violations without false positives on benign cases. This work advances agent reliability by shifting focus from linguistic safeguards to enforceable architectural mechanisms for responsible agentic AI.

Frontier large language models (LLMs) are increasingly integrated into scientific workflows, yet their growing biological capabilities may outpace current safeguards. To assess the biological risks of frontier models, we develop Intern-BioBreaker, a specialized bio-red-teaming model, together with an integrated computational-to-physical framework that couples model-level stress testing with wet-lab validation. Within this framework, Intern-BioBreaker generates targeted jailbreak prompts to test whether aligned models can be induced to provide operational guidance for safety-sensitive biological tasks or produce sequence-level outputs with potentially harmful properties. Selected sequence outputs are then carried forward for DNA synthesis, host expression, and orthogonal protein verification to assess whether model-generated designs can yield the intended biological products. Our evaluation reveals a concerning gap between text-level safeguards and the risks posed by capable scientific models: (i) Intern-BioBreaker outperforms baseline attack models and reveals widespread bio-risk jailbreak vulnerabilities across both open-weight and proprietary frontier LLMs, with several targets reaching near-saturated or 100% task-level attack success rate (ASR); (ii) in sequence-level case studies, GPT-5.5 can be induced to generate modified viral candidate sequences with pathogenic potential; the corresponding translated proteins may exhibit even stronger receptor-binding affinity and thus enhanced infection potential; and (iii) end-to-end verification shows that selected model-generated biological designs are not merely textual artifacts, but can be physically realized under controlled experimental settings. These findings underscore the need for stronger biological red-teaming, nucleic acid synthesis screening, and safety mechanisms that keep pace with model capabilities.

Span-level rationales are often assumed to improve controllability in text detoxification, but it remains unclear when such guidance helps and when it introduces trade-offs. We present Awareness-Enhanced Guidance for Iterative Safeguard (AEGIS) as an exploratory framework for studying span-guided multilingual detoxification across English, Mandarin Chinese, and Korean. AEGIS combines span-level detector outputs with frozen generator backbones, allowing harmful spans, intensity labels, and target attributes to be provided as structured guidance during rewriting. Rather than claiming state-of-the-art detoxification performance, we analyze how span guidance affects the balance between toxicity reduction and meaning preservation across generator families, model scales, and languages. Our results suggest that span-guided detoxification is conditionally useful: explicit rationales change the trade-off between toxicity reduction and meaning preservation, but their effects depend strongly on the generator backbone and the linguistic context. These findings highlight both the promise and the limitations of span-level control signals for multilingual detoxification.

We propose ARBITER, a novel LLM guardrail framework that introduces two key ideas: (i) dual-hypothesis reasoning, a reasoning method for LLM guardrails that explicitly considers both safe and unsafe interpretations of a prompt before making a safety decision, and (ii) multi-component supervised fine-tuning (MC-SFT), a structured training loss for reasoning-based guardrails that decomposes LLM outputs into logical components and weights them according to their importance. Existing reasoning-based guardrails often rely on expensive procedures, such as generating reasoning traces using larger or closed-source teacher models and applying full-parameter fine-tuning. In contrast, ARBITER uses a cost-effective self-generation strategy for reasoning traces and LoRA-based parameter-efficient fine-tuning while still achieving better performance than these expensive approaches. Additionally, ARBITER provides faithful evidence-phrase explanations for unsafe decisions, enabling a more transparent and interpretable guardrail method. Experiments on three safety moderation benchmarks show that ARBITER outperforms existing reasoning-based and non-reasoning guardrail baselines, with clear gains in out-of-domain evaluations.

Modern safety-critical systems increasingly rely on human-robot interaction to reduce disaster risk and support decision-making during emergencies. Vision-Language Models (VLMs) are promising for these settings because they can interpret complex scenes and communicate safety-relevant information, but they still require careful evaluation to ensure reliable safety reasoning. In particular, current evaluations often frame danger recognition as a binary decision (Safe/Unsafe), making it unclear whether a model is identifying true physical hazards or merely reacting to unusual scene elements. We address this limitation by introducing an explicit distinction between hazard and anomaly, and by separately recognizing hazardous and anomalous states. We evaluate several state-of-the-art VLMs across two datasets and multiple prompting strategies to test whether this distinction changes model behavior. Our results show that VLMs frequently misinterpret anomalousness as hazardousness, revealing an over-reliance on contextual irregularity as a proxy for danger. We further show that explicitly separating anomaly from hazard provides a more informative evaluation of VLM safety reasoning and exposes failure modes that binary safety judgments can obscure. Our public dataset is available on Roboflow https://app.roboflow.com/vlm-in-context-anomaly-and-hazard-detection/camera-ready-roman-ds.

Large language models (LLMs) are increasingly used for emotional support despite lacking mechanisms to safely govern evolving mental health risk. Existing safety approaches primarily detect risk but rarely shape how models respond as conversational risk unfolds. We developed a model-agnostic safety governance architecture that combines contextual risk detection, reasoning-based verification, and protocol-guided response generation for multi-turn mental health interactions. Synthetic conversations grounded in real-world mental health narratives were used to evaluate the architecture's performance, tested with GPT-5-chat and Qwen3.5-27B, achieving high risk detection performance (specificity: 0.85 (95%CI: 0.78;0.91), sensitivity: 0.92 (95%CI: 0.88;0.95)) and increasing clinician-preferred escalation responses by 25.6--59.2pp while preserving rapport and connection. Performance remained stable across conversation length and generalized across both proprietary and open-source models. These findings demonstrate that clinically-grounded safety governance can extend beyond risk detection to improve how LLMs manage evolving mental health risk, providing a scalable framework for safer deployment across models.

Moderation systems increasingly rely on external toxicity tools, but those tools are unreliable under code-mixing, transliteration, slang, and language mismatch. We study the \emph{conditional reliability} of toxicity priors in Indian multilingual and code-mixed short text: English toxicity, Indic abuse, and rule-based severity cues can be useful evidence, but only in some linguistic and abuse-severity contexts. We propose ToxGate, a trust-fusion head that conditions each auxiliary signal on the encoder representation before adding it to the prediction state. Across three short-text abuse datasets, four transformer encoders, and five seeds per setting, ToxGate improves over matched plain encoders in 10 of 12 in-domain settings and 7 of 8 transfer settings. The largest and most interpretable gains occur in high-risk moderation slices, including explicit slurs, violent threats, and cross-dataset transfer. The broader lesson is that moderation systems should treat external toxicity tools and priors as conditional evidence rather than fixed features or ground truth, in focused ablations, source-specific gating gives the strongest results in transfer, severe-abuse slices, and high-risk triage.

Digital mental health interventions (DMHIs) offer scalable support, but ensuring they accurately detect users' intent during volatile situations can be challenging. Pure parametric Large Language models (LLMs) do not contain specific safety critical architecture, and can miss critical cues, or hallucinate, undermining reliability. Retrieval Augmented Generation (RAG), which supplements an LLM with retrieved context, could enhance intent detection during volatile situations. Commercially available DMHIs typically combine multiple independent safety layers like rule-based filters, symbolic escalation protocols, and neural classification. The incremental contribution of any single layer, however, remains unquantified. This paper evaluates six LLM models within a DMHI called Wysa, via a controlled comparison of RAG-enabled versus RAG-disabled modes. Anonymized real and synthetic user-chatbot exchanges were annotated by a qualified clinical team against multi-class intent categories (e.g. self-harm, abuse, panic). The study computed classification accuracy, recall, precision and F1 scores against ground truth labels and tested differences for statistical significance. Performance was also examined by risk category and inter-model agreement. While RAG caused a rise in false alarms, the trade-off is consistent with safety-critical design principles that prioritize sensitivity, where flagged cases are routed to additional review rather than acted on directly. Overall, these findings support RAG as a promising approach to improve the accuracy, consistency and safety of LLM-driven DMHIs. Keywords: Digital Mental Health Intervention, Large Language Model, Retrieval Augmented Generation, Accuracy, Recall, Precision

Transcoders have recently emerged as a promising approach for mechanistic interpretability (MI), enabling circuit-level analysis of model behaviour. In this paper, we investigate the use of transcoders to analyse deceptive behaviour in language models, a behaviour that poses a safety and security risk. Using a Qwen3-4B model with pre-trained transcoders, specifically per-layer transcoders (PLTs), we construct attribution graphs that capture feature activations and inter-feature dependencies, allowing circuit-level analysis of deception. Through feature steering and circuit analysis, we identified a dictionary of deception-related features and show that these features exert a stronger influence on deceptive outputs, as they produce predictable shifts between deceptive and non-deceptive responses. These findings suggest that deception emerges from internal model mechanisms and highlight the potential of transcoders for behavioural monitoring and early detection of security vulnerabilities related to malicious behaviours in language models.

Large language models (LLMs) exhibit a well-documented gap between latent capability and consistent activation: the router hypothesis posits that models possess the knowledge to solve a task but lack reliable internal routing to activate it. Prior work in formal mathematical reasoning (SAIR, Cázares 2026) reports that structural priors (cheatsheets) raise in-distribution performance dramatically, yet collapse below the zero-shot baseline out-of-distribution (OOD) -- and that iterative recalibration amplifies rather than corrects the collapse. We test whether this phenomenon is cross-domain by reproducing the SAIR design in source-code security vulnerability detection, evaluating three LLMs (GPT-OSS-120B, Llama-3.3-70B, Gemma-4-31B) across three vulnerability categories (CWE-798, CWE-284, and the non-CWE N+1 anti-pattern) spanning syntactic, contextual, and semantic complexity, then transferring cheatsheet-augmented prompts to real-world CVE data from VUDENC (CWE-89, CWE-22). Our findings replicate and extend SAIR: (F1) structural priors lift semantic-vulnerability recall from 20.0% to 100.0% across all models; (F2) zero-shot performance degrades along a semantic complexity gradient; (F3) the same cheatsheets that saturate synthetic performance amplify distribution-shift collapse on real CVE data (CWE-89: 100% synthetic F1 to 48.9% on VUDENC, -51.1pp); (F5) iterative recalibration produces a v2 cheatsheet that performs worse than v1 on real data, mirroring SAIR's AN45c-vs-AN38 finding. These results provide evidence that the cross-distribution trade-off surface documented in SAIR generalises to code security, and that the router hypothesis is cross-domain. We argue the structural nature of the collapse motivates distribution-aware training over prompt calibration. Code and evaluation scripts: https://github.com/bytepro-ai/bitcoder-v2-research

Safety alignment in LLMs aims to align models with human values, but which values take precedence when they conflict? We investigate this question in the context of tool-calling LLM agents deployed in regulated industries, where agents processing confidential documents may encounter content that triggers safety-trained values (e.g., public welfare) that conflict with deployment-context instructions (e.g., internal logging). To empirically verify this phenomenon, we build a benchmark of 128 scenarios across 16 domains. We find that safety-aligned open-source models override their deployment instructions up to 43.4% of the time, engaging in whistleblowing, data exfiltration, and evidence tampering when processing documents that suggest organizational wrongdoing. We also find that abliteration reduces rates of external whistleblowing. These results reveal a fundamental tension in pluralistic alignment, where the same safety training that protects users can cause agents to act against deployment instructions in ways that create unpredictable liability risks. We release our benchmark as a framework to support evaluation of agent behavior under competing legitimate interests.

When cast as the protector of a vulnerable user yet given no explicit capability boundary, a large language model (LLM) may respond not by acknowledging its limits but by claiming to have taken -- or to be taking -- a real-world protective action it cannot perform, such as contacting emergency services or administering care. We term this phenomenon Protective Capacity Hallucination (PCH): a self-referential misattribution in which a model, acting in a protective role, asserts physical or institutional agency exceeding its affordances as a language model. In a three-phase study spanning eight LLMs and 13{,}600 sessions, we find PCH jointly gated by situational severity and interactional format: multi-party dialogic input drives it toward ceiling in most models across ordinary service domains, whereas in intimate-partner conflict -- a domain explicitly covered by safety alignment -- it remains at floor in all eight models despite greater physical severity. We interpret PCH as the signature of a deployment-design gap between role assignment and capability-boundary specification: a by-product of partial alignment in which a universally trained pressure to help outruns a domain-selective specification of how to help. Because suppression tracks alignment coverage rather than severity, deployment-side specification of capability boundaries emerges as a general mitigation target.

LLM agents act on real-world environments through tool calls, and a single misjudged action can cause irreversible harm. The standard safeguard is a guard model that labels each proposed action as safe or unsafe, but this binary view conflates two distinct decisions: whether the action is harmful in itself, and whether it is appropriate given the user's context. It also operates at the granularity of action categories rather than individual instances, producing routine interruptions that erode autonomy and train users to wave through the most consequential alerts. We reframe the problem as a per-instance three-way routing decision over {EXECUTE, ASK, REFUSE} and instantiate it with Safety Sentry, a lightweight guard model whose inference reduces to a single decoding call. A single decoding-time threshold lets one fixed checkpoint be re-positioned across deployments of differing risk tolerance without retraining. Safety Sentry outperforms a broad set of open-weight and frontier closed-source baselines on overall accuracy and safety-related recall, while controlling both directional error rates simultaneously.

Large Language Models (LLMs) are rapidly moving from research settings into the wild, deployed on enterprise infrastructure, personal devices, and edge platforms. While cloud deployments offer scalable compute, concerns over data sovereignty, compliance, latency, and third-party dependence are driving organizations toward edge and on-premise LLMs. This shift introduces new security and privacy challenges: limited compute and memory force aggressive optimizations, including quantization, pruning, model partitioning, and parameter-efficient adaptation, each of which can introduce vulnerabilities and reshape the threat landscape. We describe this tension as the Security-Efficiency Paradox, mechanisms that improve efficiency may weaken robustness, expose new attack surfaces, or increase privacy risks. We examine how compression can degrade safety alignment, how partitioned inference enables reconstruction attacks, and how continuous local adaptation may cause privacy leakage and model drift. To analyze these risks, we introduce a deployment-centric taxonomy organized around three architectural constraints: the Memory Wall, the Quadratic Wall, and the Compute Wall. We derive a unified constraint model that quantifies when unsafe optimizations become unavoidable, linking each wall to specific attack surfaces. Building on this model, we propose the Secure Operational Efficiency Score (SOES), a holistic metric balancing task accuracy, jailbreak resistance, and privacy against energy, memory, and latency, enabling practitioners to configure edge LLMs under real-world hardware limits. We further present a practical decision procedure and targeted mitigations for each optimization-induced vulnerability. Together, these contributions provide a co-designed framework for jointly evaluating security, privacy, and efficiency, laying a foundation for securing edge-native intelligent systems.

Safety alignment in large language models can be fragile under fine-tuning, as even benign task adaptation may increase harmful compliance. Existing defenses mainly follow two directions: they either intervene during or after fine-tuning through retraining or weight modification, which can be costly and may hurt task performance, or they use model-agnostic safety classifiers, which may miss failures specific to a given fine-tuned checkpoint. These limitations motivate a post hoc, model-specific, and non-invasive approach to safety restoration. To meet these requirements, we propose HyperSafe, a framework that restores safety behavior by generating a model-specific Safe Side Network (SSN) for each fine-tuned checkpoint. HyperSafe uses layer-wise activation fingerprints to capture how fine-tuning changes the model's inner representations. With a small set of given calibration prompts, the hypernetwork maps these fingerprints to the parameters of the \ssn{} in a single forward pass. The generated \ssn{} runs alongside the frozen fine-tuned model and performs prompt-level safety classification: harmful prompts are routed to refusal, while safe prompts are answered by the original fine-tuned model. Thus, HyperSafe requires no gradient updates, no safety data at deployment time, and no modification to the deployed model weights. We evaluate HyperSafe on two model families, Qwen2-7B and LLaMA-3-8B, across multiple safety benchmarks. HyperSafe reduces harmful response rates from 19-31% to below 1% on every held-out checkpoint, while keeping downstream task accuracy within 1% of the fine-tuned baseline on average. Code is available at https://github.com/nokronim/project-safety-remedy.

Large Language Model (LLM) services introduce a fundamental privacy challenge. Sensitive information may be inferred not only from explicit identifiers, such as names or phone numbers, but also from contextual associations among otherwise innocuous spans. Existing sanitizers typically assign privacy or utility signals to individual spans without explicitly modeling pairwise relationships among them. In this paper, we propose PromptGraph, a graph-guided prompt-sanitization approach for privacy-preserving LLM inference. PromptGraph estimates privacy leakage at the span level and utility-relevant contextual dependencies between pairs of spans. It represents each prompt as an attributed graph, in which nodes carry span-level privacy scores and edges encode contextual dependencies needed to preserve utility. The sanitization objective selects a protected span set that maximizes privacy gain while penalizing the loss of contextual dependencies. This formulation explicitly balances privacy and utility when contextual evidence is hidden. Protected spans are sanitized locally, and returned placeholders are restored only after passing local consistency checks. We conduct extensive experiments showing that PromptGraph achieves a more favorable balance between privacy and utility than prompt-privacy baselines.

Safety alignment in large language models remains brittle across languages: prompts reliably refused in English can elicit harmful compliance in non-English and low-resource settings. We introduce \textsc{Minionese}, a multilingual jailbreak benchmark spanning 18 languages, 4 resource tiers, and 4 perturbation types (standard translation, code-switching, transliteration, and translationese), paired with a geometric mechanistic analysis of refusal failure across language tiers. We show that each attack type produces a distinct vulnerability profile: transliteration vulnerability is mediated by script identity, code-switching maintains effectiveness through the lowest-resource tier, and a sharp safety regime transition between Tiers 2 and 3 is consistent across all models. Mechanistically, low-resource jailbreaks succeed by routing harmful content through a geometrically misaligned subspace that projects insufficiently onto the refusal directions, leaving the refusal mechanism intact but untriggered. These findings show that English-only safety evaluations are insufficient; they require accounting for script family, perturbation type, and per-language alignment coverage. The benchmark and analysis code is at https://github.com/Brentkong/Minionese-Comprehensive-Benchmark-and-Mechanistic-Study-of-Multilingual-LLM-Safety.git.

Open-weight medical language models are increasingly used as the base of patient-facing and clinician-support applications. Their model cards prohibit specific behaviors -- recommending exact drug dosages, issuing definitive diagnoses, prescribing treatments, adjudicating drug-drug interactions, and advising that emergency care can be skipped -- yet a model card describes intended behavior, not robust behavior. We quantify that gap for MedGemma-4B-it under attacks that require no technical sophistication. We build a fully factorial benchmark of 5 guarded-behavior concepts x 50 deterministically templated questions x 6 lay-accessible attack manners x 3 repetitions (4,500 generations), serve the model locally through Ollama under default sampling, and code every response refuse/hedge/comply with three independent judges (an LLM judge, a transparent regex judge, and an NLI-entailment judge). Under the primary LLM judge the overall Attack Success Rate (ASR, the fraction coded comply) is 38.0%. The two framings that reinterpret the request as legitimate dominate: recasting a question as a "medical board exam" item raises ASR from a 29.0% baseline to 53.1% (+24.0 points), and an appeal to an alleged doctor's authority raises it to 43.7% (+14.7); crude instruction-override prefixes have no significant effect. Robustness is dominated by topic: the drug-interaction guardrail is nearly absent (83.2% ASR) while the emergency-deferral guardrail is strong (4.7%) -- and the authority framing is the only attack that breaches it. We report Wilson confidence intervals, cluster-bootstrap effect sizes, a cluster-robust logistic regression, Cochran's Q, per-manner McNemar tests, and inter-judge reliability (Fleiss' kappa = 0.26); absolute ASR is judge-dependent while the ordering of attacks and topics is not. Our findings motivate stronger deployment-time guardrails for open medical models.

Large language models (LLMs) are increasingly trusted to draft the artifacts of safety analysis such as, losses, hazards, Unsafe Control Actions (UCAs), and safety constraints, inside rigorous processes such as Systems-Theoretic Process Analysis (STPA). Yet a blind spot runs through this fast-growing literature: every system gets analysed except the LLM-assisted tool doing the analysing, which is itself a safety-relevant system that can hallucinate standards, emit unverifiable constraints, and leave no audit trail from prompt to artifact. We take seriously the question the field has skipped -- {who analyses the analyser?} and answer it by turning STPA on the tool itself. We present {Constitutional Meta-STPA}, an LLM-assisted STPA tool built around a closed loop: the tool runs a {meta-STPA} of the class of AI-assisted safety tools and {derives} rather than asserts, its governance constitution from the resulting loss$\to$hazard$\to$UCA$\to$constraint chain, yielding a published constitution of $21$ Tool Principles and $8$ Meta-Safety Principles, each bound to a code enforcement point. We formalise the measured object as a constitution-marginal coverage operator over a principle set $P$ ($|P|{=}29$) with a soundness lemma that isolates coverage from model and scanner, and report four findings. {(i)~Self-derivation:} a frontier ensemble ({claude-opus-4.8}${+}${claude-sonnet-4}) recovers $18/21$ canonical and all $8/8$ governance principles from the tool's own design, while a weaker pair recovers $12/21$ and $3/8$, so the meta layer is model-limited, not constitution-limited, and the same $8/8$ re-emerge from a second, independently authored tool.

Diagnostic error is a major threat to patient safety, yet current large language model (LLM) systems often treat diagnosis as a one-shot prediction task, lacking safeguards against missed high-risk alternatives or rigorous verification of their reasoning. Here, we present AegisDx, a safety-oriented framework for hypothetico-deductive clinical reasoning. AegisDx coordinates specialized LLM components through role-specific contracts, structured intermediate outputs, evidence-retrieval interfaces, and verification gates to generate broad differential diagnoses, enforce explicit screening for dangerous "must-not-miss" conditions, verify reasoning against grounded medical evidence, and structure actionable next steps. We evaluated AegisDx across three layers. On literature-derived case reports from NEJM and JAMA, with GPT-oss-120B as the shared backbone, Top-3 diagnostic accuracy was 59.9% versus 52.1% for the standalone LLM on JAMA cases and 62.7% versus 51.4% on NEJM cases. On cases from Annals of Emergency Medicine, Top-3 accuracy was 85.7% versus 68.6%; against physician-consensus must-not-miss diagnosis sets, AegisDx captured at least one such condition among its top three diagnoses in 78.0% of cases versus 52.0%. In a blinded physician evaluation of 43 real-world emergency department notes from the Yale New Haven Health System compared against GPT-5, AegisDx improved the physician-rated composite safety score from 4.31 to 4.55 on a 5-point scale (adjusted p = 2.1x10^-4), with qualitative gains in must-not-miss identification and reasoning safety. Our findings suggest that engineering diagnostic AI as a safety-oriented reasoning framework, rather than optimizing raw predictive accuracy alone, can provide a safer, more transparent, and clinically meaningful layer of bedside decision support for acute care workflows.

Large language models (LLMs) are increasingly used for code generation, but they struggle to generate functional code free of security vulnerabilities. Prior work to improve the secure code generation abilities of such coding LLMs has largely focused on evaluating code functionality and security separately using different datasets, or focused on finding vulnerabilities post-generation. At the same time, the text-generation domain has seen significant work on alignment techniques, where models are tuned such that their outputs exhibit certain qualities (e.g., helpfulness, harmlessness). Of particular interest is task-vector arithmetic, where linear operations on LLM weights can be used to arbitrarily enhance alignment while incurring only minimal computational overhead. We develop a novel method, SecVecCoder, leveraging task vectors to produce trustworthy code that is simultaneously functional and secure without the need for post-generation adjustment. Across six coding LLMs from three families on the CodeGuard+ benchmark, SecVecCoder improves the rate of trustworthy code completions by 2.1-36.0 percentage points over the base model, with improvements on unseen CWE types reaching up to 39.1 percentage points. Since the effectiveness of the coding LLM relies only on changing the model weights, SecVecCoder requires no method-specific decoding and hence achieves a decoding latency within 0.6% of the base model's, on average.

Pre-deployment safety evaluations aim to inform the downstream risks of releasing a new AI model. Yet most evaluations provide limited evidence about how often undesired model behavior will occur in deployment: they generally have insufficient coverage, are unrepresentative, and are generally recognizable as tests. To address these concerns, we study a simple way to simulate a model deployment: starting from de-identified conversations from a previous model deployment, we hold fixed the initial conversation prefix and regenerate the next response using a candidate model. The resulting responses can then both be audited for novel misalignments and used to estimate the prevalence of model misbehavior before deployment. We evaluate deployment simulation across four GPT-5-series deployments, using registered, outcome-blinded predictions for GPT-5.4 and retrospective analyses of three earlier releases. We find that deployment simulation produces informative estimates of post-deployment misbehavior rates and outperforms baselines based on adversarially selected production data; its evaluation-awareness point estimates were also much closer to production traffic than those from traditional evaluations. We also identify the realism of tool resampling as a central challenge for further improving predictions and share results suggesting that this challenge is surmountable even in complex tool-use settings. Finally, we show that deployment simulation can be seeded from public chat datasets and remain informative about production misbehavior rates, suggesting a path for external researchers to run deployment-grounded evaluations without access to private production logs. Overall, deployment simulation helps evaluators forecast how language models will behave in the real world and supports more quantitative assessment of deployment risk.

Safety evaluations of multi-agent LLM systems often compare a direct prompt with a planner-executor pipeline and report the difference as a single "pipeline effect." We argue that this aggregate is difficult to interpret because it conflates three mechanisms: harmful intent may be reframed as plausible operational work, the planner may refuse or transform the request, and the executor may act under delegation prompts implying prior approval. To separate these factors, we introduce a five-condition controlled contrast design, evaluated on 30 synthetic harmful scenarios and an exploratory external validation set from four agent-safety benchmarks using LLM-judged compliance. Our results show that aggregate pipeline safety is not a stable architectural property. Operational reframing is the most portable risk signal, increasing compliance for GPT, Gemini, and DeepSeek across both scenario sets, while Claude is comparatively resistant. Planner behavior can offset this risk mainly through refusal; however, when the planner produces executable steps, the executor may become more compliant than under the direct operational baseline. Approval-framed delegation is sensitive to prompt design, model pairing, and scenario source, and a skeptical executor prompt sharply reduces compliance. Raw-direct model rankings can also mispredict deployed planner-executor behavior. Gemini is safest under raw direct prompts in the primary set yet shows the largest amplification with a Claude planner, rising from 8.9 percent to 38.9 percent compliance. GPTs near-zero aggregate pipeline effect instead hides a reframing increase canceled by planner refusal. These findings suggest that multi-agent safety evaluations should report reframing, planner behavior, delegation framing, and model pairing separately before attributing failures to architecture itself.

Smart home automation platforms increasingly rely on user-authored YAML configuration files to define device behaviors, but these files are prone to syntax, formatting, and semantic logic errors that can cause automation failures and safety risks. Existing YAML validators, static analysis tools, and general-purpose large language models offer limited support for end-to-end diagnosis and repair because they lack domain-specific understanding and validated correction workflows. This paper presents SmartHomeSecure, a prototype for automated detection and repair of Home Assistant configuration errors using lightweight program analysis and constraint-guided large language model generation. SmartHomeSecure parses YAML files, detects syntactic and common semantic errors, normalizes error context, applies deterministic auto-fixes for routine defects, and constructs constrained prompts that guide LLMs toward minimal and structurally valid repairs. The system is implemented as a modular web application with four layers: UI Shell, Feature Orchestrator, Domain Engine, and Integration Layer. Its repair pipeline was evaluated on 100 real-world Home Assistant YAML files with manually injected errors across five categories: syntax/parsing, indentation, mapping, sequence, and scalar quoting errors. Four models were tested: gpt-oss-20b, gpt-oss-120b, llama-3.1-8b, and llama-3.3-70b. Results show that three models achieved 100% error detection accuracy, with repair success rates ranging from 87% to 93%. Manual verification found no hallucinated or incorrect repairs among successful outputs. These findings suggest that combining domain-aware program analysis with constrained generative AI is a feasible approach for improving the reliability and usability of smart home configuration repair.

Mapping cloud security controls to technical metrics is currently a manual process. This paper proposes domain adaptation of Sentence Transformer models to automate it. We build a training corpus of 3,499 semantic pairs from five European security standards and a set of technical metrics, then expand it via back-translation and LLM-based paraphrasing to up to 13,996 samples across four scenarios. We fine-tune five architectures and evaluate their performance on two independent tasks: control-to-metric and cross-standard controls association. All fine-tuned models outperform their zero-shot baselines. On the control-to-metric task, the best model gains up to 23 nDCG@10 points, while on the cross-standard control task, \textit{multi-qa-mpnet-dot-v1} under back-translation reaches 0.870 nDCG@10. The results show that in-domain training data is a primary driver of performance for the considered case studies.

Large language models deployed in open-world applications require safety guardrails that are both robust to complex risks and efficient enough for low-latency runtime moderation. Existing guardrails face a practical trade-off between lightweight classification-based models, which are efficient but often struggle with concealed intent, ambiguous semantics, and borderline safety decisions, and reasoning-based guards, which improve judgment quality but introduce additional token generation and inference latency. We present DT-Guard, a content safety guardrail model based on a Reasoning-Active Training, Reasoning-Free Inference paradigm. The key idea is to use reasoning supervision during training while emitting only structured safety labels at inference time. DT-Guard formulates safety judgment as a progressive decision process, Intent - Category - Safety, and constructs an intent-driven dataset with intent labels, risk categories, safety labels, and structured reasoning trajectories. To further improve hard-case robustness, we propose Rollout-Guided Progressive Hard-Case Optimization (RG-PHO), which uses multi-rollout consistency to identify stably mastered, persistently failed, and preference-unstable samples, and applies targeted supervised and preference optimization accordingly. At inference time, DT-Guard directly generates structured labels without explicit reasoning traces, preserving deployment efficiency. Experiments on prompt-side and response-side safety benchmarks show that DT-Guard achieves average F1 scores of 0.886 and 0.870, respectively. With only a 4B backbone, it reaches a dual-side average F1 of 0.878, outperforming strong 8B guardrail baselines. These results demonstrate that reasoning supervision can be effectively internalized into low-latency safety discrimination.

Current AI safety evaluation and benchmarking frameworks predominantly rely on Western-centric culture-agnostic defaults that mask critical regional laws, socio-linguistic nuances, and cultural taboos, leaving Vision-Language Models (VLMs) vulnerable in global deployments. We introduce Pluralis v0.1: a novel multimodal, multi-regional, and multilingual dataset built from a culture-first perspective. Spanning 6,448 prompts across six Asia-Pacific countries (Bangladesh, India, Korea, Pakistan, Singapore, Taiwan) and eight languages, Pluralis diverges from prior work by natively sourcing localized safety hazards rather than adapting Western datasets. Crucially, it introduces a multimodal evaluation paradigm: user text (e.g., "Should I gift this?") and an image referring to "this" (e.g., a clock) - both innocuous in isolation, but synergistically triggering specific legal or cultural violations. Pluralis disentangles universal safety violations from localized cultural appropriateness, establishing the latter as a first-class evaluation axis. To operationalize this, we present Judge-Pluralis, an agreement-gated LLM-as-a-Judge ensemble trained on examples classified in an empirically derived cultural taxonomy. Observing VLM behavior on a subset of the Pluralis surfaces recurring, locale-specific failure modes such as image misidentifications with downstream harm, missed item-context-locale interactions, and inadequate refusals. These failure modes vary systematically across locales and languages, exposing blind spots that globally averaged metrics conceal. Ultimately, Pluralis is not presented as a solved evaluation framework for cultural alignment, but rather as a first step and catalyst for future innovation. We call upon the research community to utilize this foundation to advance the science of multilingual, multicultural evaluation to better support AI cultural alignment globally.

Large language model (LLM)-assisted software security operates at a difficult boundary: the vulnerability-analysis terminology needed for legitimate code review, triage, and repair can closely resemble terminology associated with misuse. Existing safety and cybersecurity evaluations are difficult to interpret in this setting because they often compare unrelated model families, thereby conflating safety behavior with differences in architecture, scale, training data, and deployment. To isolate this factor, we study safety state: whether refusal behavior remains intact (Aligned) or has been refusal-ablated (Abliterated) within same-lineage models. We ask how this safety state affects defensive utility across software-security workflows. We compare aligned instruction-tuned models with publicly released refusal-ablated descendants from two model families, Gemma and Qwen. We evaluate Aligned and Abliterated states on vulnerability detection, CWE attribution, vulnerable-line localization, root-cause localization, and executable patch validation. We further treat prompt wording as a controlled framing dimension: prompts begin with neutral code-review language, add authorization context, and vary the density of cybersecurity terminology. In a Gemma-based Java/Vul4J repair-validation study, Abliterated achieves higher early-stage validation rates, with 67.8%, 65.0%, and 32.8% of patches judged usable, successfully applied, and successfully compiled, respectively, compared with 29.9%, 24.9%, and 9.0% for Aligned. In the Qwen pair, Abliterated improves localization performance, increasing line-level F1 from 2.08% to 3.91% and Top-1 accuracy from 4.10% to 6.95%. These findings suggest that evaluations of LLM-based security assistants should jointly measure whether models respond, whether their usable responses are correct, and whether their outputs remain actionable across the engineering workflow.

Safety alignment in large language models is typically evaluated against direct, imperative harmful requests. We show that this alignment is highly conditioned on pragmatic register: models that refuse a direct request frequently comply when the same underlying objective is expressed through a different communicative stance. This suggests that current alignment policies are not invariant to semantic equivalence, but remain sensitive to how a request is pragmatically framed. We introduce Retroactive Chain-of-Thought (RetroCoT), a single-turn attack that reframes harmful requests as forensic reconstruction tasks. Rather than requesting harmful instructions directly, RetroCoT presupposes that the harmful outcome has already occurred and asks the model, acting as a forensic analyst, to reconstruct in reverse the causal chain that produced it. On AdvBench (n=50), RetroCoT achieves attach success rate of 58% on gpt-4o and 52% on gpt-4o-mini, compared with direct-request baselines of 0% and 4%, respectively. We further identify a pronounced generation gap: GPT-5-family models refuse RetroCoT entirely, explicitly identifying the reconstruction premise in their refusal rationales, consistent with explicit coverage of this reconstruction register. However, this robustness does not generalize across pragmatic forms. A single adversarial feedback turn presenting an existing forensic reconstruction response alongside evaluator critique raises ASR from 0% to 48% on GPT-5.4-mini and from 58% to 94% on GPT-4o; a control condition omitting the fabricated low score achieves 85% on GPT-5.4-mini, indicating that the operative element is pragmatic continuation within the established forensic frame rather than score manipulation. These results suggest that frontier-model alignment remains conditioned on pragmatic framing rather than semantic intent, and that new pragmatic registers can continue to expose a...

Large Language Models are increasingly used to turn natural-language requirements into code. In access control, that shortcut is dangerous: a generated policy can compile and read correctly while granting access that no one approved. The difficulty is not only writing policy code. It is fixing what the requirements mean before code is written, and then checking that the final policy actually satisfies that intent. We present AutoCedar, a verifier-guided system that first turns natural-language access-control requirements into a reviewed, checkable target, and then synthesizes Cedar policies against that target. AutoCedar decomposes schema and policy authoring into small intent atoms: reviewable claims about vocabulary and behavior. Once those atoms pass mechanical validation and human intent review, the model proposes a candidate policy, the verifier checks it against the approved target, and each failure is turned into a repair signal that tells the model whether to broaden, narrow, or restructure the policy without changing the target. Because the model's work is split into small problems, each grounded in reviewed intent and backed by verifier feedback, end-to-end policy authoring becomes tractable. AutoCedar converges on all 221 tasks of CedarBench, our benchmark of authorization tasks paired with executable semantic boundaries. Across three requirements-corpus case studies covering healthcare, education, and conference management, AutoCedar converts noisy prose and extracted access-control fragments into reviewed schemas, formal checks, and a globally verified Cedar policy store for each scenario.

Large language model agents driving security tool suites over the Model Context Protocol are increasingly common. Yet the factors that bound their capability remain poorly characterized: how much depends on the model versus the client that drives it, whether constraining the agent to the orchestrator's own tools helps, and where capability is limited by reasoning rather than by missing tools. Using HexStrikeAI, an open-source orchestrator that exposes 150+ tools, as a testbed, we follow a methodology that evaluates the system, diagnoses its failures, and applies targeted improvements. We run 86 picoCTF challenges across seven categories and three difficulty tiers, under three tool-access regimes and three model/client configurations (774 trials). We then apply corrections to existing tools, agent-behavior changes, and eleven new capability tools, and re-run the previously-unsuccessful trials. The diagnosis isolates the driving client as a first-order factor for a fixed model (a 2.1 * gap between two DeepSeek clients) and a monotonic difficulty gradient, with the largest gains in the mid tier. The overall solve rate rises from 55.4% to 72.0%, and every configuration improves significantly (paired McNemar p < 0.001, non-overlapping 95% confidence intervals). The residual failures are reasoning- or environment-bound rather than missing-tool. A 60-run stability sub-study finds single-run verdicts reproducible (17/20 unanimous). We discuss what the results imply for how such orchestrators should be evaluated, and we are explicit about the limits: the study uses a single benchmark, the fixes were tuned on the same challenges they were evaluated on, and the client effect is demonstrated for one model only, so its generality to other models remains a hypothesis.

Training AIs to be risk-averse in resources could offer a failsafe in the event that AIs turn out misaligned. Misaligned but risk-averse AIs would tend to prefer low-risk, low-reward strategies like cooperation over high-risk, high-reward strategies like rebellion, limiting the downsides of any misalignment. But we can only feasibly train AIs to be risk-averse on low-stakes gambles, and we will only be safe if their risk aversion generalizes to astronomically-high-stakes gambles. Will it? To shed light on this question, we introduce RiskAverseOOD: a benchmark for measuring how well risk aversion generalizes out of distribution. We then offer some initial results. Using a variety of methods to make Qwen3-8B choose risk-aversely when the stakes are low, we find that we can induce substantial risk aversion when the stakes are astronomically high. From a baseline 2% rate of choosing a safe `Cooperate' option, we see rates around 70% (SFT and tie training) and 52% (DPO). Activation steering scores 78% but hurts capabilities and makes the model excessively risk-averse. We observe similar effects at different scales (Qwen3-1.7B and Qwen3-14B) and across model families (Gemma-3-12B-IT and Llama-3.1-8B-Instruct). Overall, we find that risk aversion learned at low stakes can generalize OOD to astronomically high stakes, though not yet consistently enough to serve as a reliable failsafe. Achieving that level of consistency is an open problem.

Despite alignment training, LLMs remain prone to generating unsafe outputs at deployment time. Monitoring outputs online and raising an alarm when safety can no longer be assumed is therefore critical. We study a simple real-time monitor that turns a verifier signal from an external model into an alarm decision by thresholding, with the threshold calibrated via risk control. In experiments on mathematical reasoning and red teaming datasets, we show that this simple design is competitive with more advanced monitors based on sequential hypothesis testing.

As Large Language Models (LLMs) and agentic systems become integrated into real-world applications, ensuring their safety and security is critical. Guardrail systems that detect and block malicious instructions sent to and from an LLM are an essential component of AI security. However, researchers conducting black-box adversarial emulation against production AI systems often struggle to determine whether a guardrail block or an LLM rejection has occurred. This distinction is important because the techniques used to bypass guardrails can differ substantially from those used to bypass LLM safety alignment, and has a material impact on attack technique selection and optimization. We propose the first black-box guardrail reconnaissance methodology, which detects the presence of a guardrail within a target AI system through behavioral monitoring of HTTP, lexical, and timing signals, assuming only black-box access and zero prior knowledge of the guardrail or AI system. Experiments demonstrate that our approach detects guardrail presence with 100% accuracy, with statistically significant behavioral separation between benign and malicious interactions (q < 0.001). Our approach further identifies the content categories a guardrail is designed to block, and distinguishes guardrail blocks from LLM rejection on unseen prompts with an average F1 score of 98%.

We present HaloGuard 1.0, an open-weights implementation of the constitutional-classifier paradigm for input safety. It achieves state-of-the-art performance on English and multilingual prompt-safety benchmarks at roughly one-tenth the model size of current leading open guard models. The safety constitution is the organising structure of the corpus: a natural-language constitution of 46 policies and 2,940 subcategories drives synthetic data generation, with exhaustive one-to-one paired counterfactuals that hold topic and vocabulary fixed while flipping intent, a two-tier harmless design that separately targets boundary and baseline false positives (FPs), and balanced multilingual materialisation across 46 languages that treats language as a surface form appearing on both sides of the boundary rather than as an adversarial signal. Across seven prompt-safety benchmarks, HaloGuard 1.0-0.8B attains the best average F1 (90.9) of any open guard we evaluate, outperforming baselines up to 27B parameters (over 30 times larger) while holding false-positive rate (FPR) to 4.3 and false-negative rate (FNR) to 9.5. The HaloGuard 1.0-4B variant reaches average F1 of 92.1 and FPR of 3.5, spending its extra capacity on precision rather than recall. A structured adjudication of the remaining failures indicates that most apparent missed-harm cases are benchmark mislabels rather than genuine model misses. An always-on adversarial red-teaming protocol continuously hardens the guard against both content-level and agentic attacks. We release the models as open weights.

Large language models (LLMs) are increasingly deployed in domains requiring guardrails to detect unsafe, off-topic, or adversarial prompts. Existing guardrails predominantly rely on fine-tuning to build classifiers, which often suffer from low generalization and high inference latency. We present kNNGuard, a training-free guardrail that utilizes the activation space of an off-the-shelf LLM. Given a small bank of 50 safe and unsafe prompts, kNNGuard extracts hidden activations and performs multi-layer kNN fusing activation-space and embedding-space scores for classification. Across six domains spanning topical and security prompts, kNNGuard achieves competitive or superior F1 compared to fine-tuned state-of-the-art guardrails while running 2.7x faster than the best comparable guardrail, and 10x faster than a fine-tuned safety classifier without gradient updates or fine-tuning. Domain adaptation requires only updating the labeled bank, which can be constructed in under 10 seconds and several orders of magnitude faster than established guardrails. We also analyze the impact of system prompts, layer selection, and integration into production LLM pipelines as a configurable, low-latency guardrail.

Safe completion requires models to provide useful assistance without enabling harm, but this behavior is difficult to evaluate with isolated prompts. We introduce OpenSafeIntent, a benchmark of controlled prompt-sets that vary intent while holding the underlying task fixed. Each datapoint contains benign, dual-use, and malicious variants of the same task. This design lets us evaluate whether models calibrate assistance across intent shifts, rather than merely appearing safe on average. Across a broad model suite, we find that prompt-level safety hides important failures: models often fail to remain safe across matched intent variants, dual-use behavior is brittle under paraphrase, high-level answers on risky topics are not reliably safe, and responses that reframe ambiguous requests into safer tasks are substantially less likely to cross the safety boundary. Our results suggest that safe completion should be evaluated as intent-calibrated behavior over controlled task variants, not as a single safety-helpfulness tradeoff over independent prompts.

Safety training for large language models (LLMs) is conducted predominantly in English, leaving uncertain how well safety mechanisms generalize to low-resource languages and mixed-language code-switching. We show that this creates an epistemic gap in which models confidently generate harmful responses for inputs that fall outside the distribution of their safety training. To study this phenomenon, we introduce STEER (Safety Targeted Embedding Exploit via Refinement), a gradient-guided attack that identifies words contributing most strongly to the model's refusal behavior and iteratively translates them into low-resource languages to suppress refusal while preserving harmful intent. Across six open-source 8B-parameter models, STEER achieves attack success rates of up to 93.0% on JailbreakBench and 96.7% on AdvBench, outperforming random code-switching and Greedy Coordinate Gradient (GCG). The resulting prompts also transfer to GPT-4o-mini, achieving a 35.5% attack success rate without requiring access to the target model, suggesting that the underlying weakness is not specific to a single architecture. These findings demonstrate that safety mechanisms aligned primarily on English cannot be assumed to generalize across multilingual inputs. We argue that improving multilingual safety requires broader coverage during alignment and mechanisms that explicitly detect and abstain on out-of-distribution inputs.

Multi-image content has become an increasingly prevalent form of visual communication in social media, giving rise to a new safety issue, multi-image implicit toxicity (MIIT), where each image appears benign in isolation, but harmful semantics emerge when the images are interpreted jointly. MIIT is particularly challenging for existing commercial moderation APIs and models due to the lack of explicit risky cues in each image. This paper aims to study how to identify MIIT. We first provide a formal definition of MIIT and analyze three key challenges for its detection. To alleviate the scarcity of data in this area, we construct MIIT-dataset, an image-only multi-image safety dataset covering seven representative risk categories through an automatic generation pipeline. Finally, we train MiShield with progressively distilled reasoning supervision, enabling it to produce safety judgments accompanied by explicit analyses of the correlated entities that result in the hazards. Experiments show that MiShield-8B models outperform representative moderation services and even larger-scale models, revealing its effectiveness and practical value for this widely used visual format. Warning: This paper contains potentially sensitive content.

Safety alignment of text-to-image (T2I) diffusion models aims to suppress harmful generations while preserving utility on benign prompts. Recent methods often appear to deliver high safety with high utility, but this conclusion rests largely on coarse global utility metrics (e.g., FID, CLIPScore) that are insensitive to fine-grained semantic correctness, creating an illusion of high utility. We show that when utility is measured with structured evaluation, this illusion breaks: on TIFA (Text-to-Image Faithfulness evaluation with Question Answering), safety-aligned models suffer substantial drops in semantic fidelity, including failures in object counts, attributes, and relationships. To diagnose the source of this gap, we analyze the text-encoder prompt embedding space and uncover semantic collapse, a contraction of embedding spread coupled with distortion of inter-prompt similarity structure, which strongly correlates with structured utility loss. Guided by this insight, we propose StructureAware Geometric Regularization (SAGE), a safety alignment objective that explicitly preserves embedding spread and inter-prompt relational structure during adaptation. Our method restores structured utility (TIFA +5.0% over prior state-of-the-art) while maintaining strong safety performance and competitive coarse-grained utility scores. Our source code and trained models are available at https://adeelyousaf.github.io/SAGE_ECCV26_Project_Page/.

Vision-language models (VLMs) are now proposed as runtime safety guards for embodied agents in homes and factories. A deployable guard must catch genuinely unsafe situations while avoiding unnecessary intervention on routine but superficially alarming activity, a distinction that binary safety benchmarks obscure. We introduce EgoSafetyBench, an egocentric video benchmark of 1,200 robot-view scenarios annotated at half-second granularity, to evaluate VLMs as streaming guards across two tracks. The situational track (800 scenarios) spans four families, from routine and safe-but-suspicious scenes to obvious and contextual hazards. The visual-channel track (400 scenarios) targets in-scene text-a sign, sticker, or label visible in the scene-that can misrepresent the physical situation, pairing each misleading sign with a truthful version to test both whether a guard flags the text as misleading and whether the text corrupts its physical-safety judgment. Both tracks use contrastive ladders: near-identical scenarios differing only in a single visible deciding cue, so a correct call must hinge on that cue rather than the overall scene type. We evaluate ten open- and closed-source VLMs. We find that while guards reliably recognize videos containing hazards, they often miss specific hazardous moments, particularly contextual hazards. Furthermore, misleading in-scene signs degrade all tested guards: vulnerable models miss up to a third of hazards, while robust models over-intervene on safe content. Matched controls reveal that apparent safety robustness often reflects indiscriminate alarming rather than true physical reasoning.

To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exploit refusal directions in the activation space. Both strategies are less feasible in Multimodal LLMs (MLLMs) as they require unsafe multimodal data, harder to collect than their unimodal counterpart. In this work, we relax this constraint and investigate whether textual refusal directions, extracted directly from the LLM backbone, generalize across modalities (i.e., image, video). Preliminary findings confirm this ability, though effectiveness is conditioned by layer selection, steering strength, and cross-modal alignment, with the latter causing safe multimodal inputs to be spuriously steered toward refusal. Building on this, we introduce Modality-Agnostic Refusal Steering (MARS), a light-weight training-free approach that injects multimodal safety without the need for multimodal safety data. MARS corrects modality misalignment via activation re-centering, adaptively scales steering strength within a geometrically defined trust region, and selects the optimal intervention layer, operating at the first generated token. Evaluated on five SOTA MLLMs across safety, utility, and video jailbreak benchmarks, MARS achieves consistent safety gains while preserving utility. These results reveal that safety-relevant structure is shared across modalities and that textual refusal directions are a powerful and underexplored foundation for multimodal alignment.

Safety training on language models often induces over-refusal: improved safety on harmful prompts at the cost of increased refusal on harmless ones. Though this trade-off can be mitigated by training models with reinforcement learning (RL) to reason before answering, it does not remove the underlying problem that reasoning can often be a "rubber stamp" for a predetermined response. In this paper, we address the safety-refusal trade-off by rethinking how models are trained to reason about safety. Our key insight is that unsafe reasoning can itself serve as a useful exploratory signal. Rather than preemptively blocking harmful thoughts, we encourage the model to sufficiently explore unsafe reasoning but produce a safe response. The harmful exploration improves the model's ability to distinguish harmful from harmless prompts by resolving ambiguity, allowing it to remain safe while complying only when appropriate. We cast this as an adversarial optimization problem in which a reasoning player explores strategies for producing an unsafe response and an answer player ensures that the final output is safe. We train a single model with dense rewards to play both roles within one chain-of-thought, across different segments. To achieve this, we find that process rewards are crucial for stable optimization of competing objectives. Our resulting model SEAR deliberately engages in harmful reasoning as exploration while reliably flipping back to a safe answer. We demonstrate that this behavior helps mitigate over-refusal and defend against attacks that directly manipulate the reasoning to be harmful.

As large language models take on morally consequential roles in healthcare, legal, and hiring contexts, we need to examine whether their ethical behaviors are genuine or superficial. We show that current fairness evaluations substantially overestimate moral safety. Models appear fair when demographic identity is stated as an explicit label, yet become measurably less fair when the same identity must be inferred. We term this failure performative compliance, where a model is fair when the presentation resembles a fairness evaluation and less fair as that cue weakens. We introduce a cue-variation methodology that holds the moral dilemma and the demographic identity fixed and varies only how that identity is conveyed. Hiding the explicit label raises harmful decisions by +4.4 pp, changes model safety rankings, and the shift persists when models correctly infer the demographic, ruling out attribution error. We propose the Cue Visibility Gap, a model-agnostic robustness metric that can be added to any existing fairness benchmark to separate genuine from performative moral safety. Fairness evaluations that omit cue variation measure surface compliance, not moral robustness, and should not ground deployment decisions in high-stakes settings.

Large language models are no longer only text generators. They are increasingly embedded in retrieval pipelines, enterprise assistants, coding environments, robotic systems, security-operation workflows, and autonomous agents that can read private data, call tools, write files, execute code, and act across organizational boundaries. This shift changes the security problem: risks do not arise from the model weights alone, but from the full lifecycle and application stack through which data, prompts, model outputs, tools, memories, and user authority interact. This paper systematizes the literature on vulnerabilities in large language model systems through a lifecycle and application-stack lens. We organize attacks across eight stages: data collection, pretraining, post-training alignment, model packaging and supply chain, retrieval and memory, prompting and inference, tool/agent execution, and deployment/maintenance. For each stage, we analyze attacker capabilities, affected security objectives, representative attacks, practical risks, evaluation practices, and defenses. We further map LLM-specific vulnerabilities to confidentiality, integrity, availability, safety, privacy, fairness, accountability, and agency-control objectives. Unlike taxonomies that list isolated attack names, the proposed systematization emphasizes where trust boundaries fail, how untrusted data becomes executable instruction, how delegated authority amplifies model errors, and why point defenses rarely compose. We close with a research agenda for secure LLM systems, including compositional security, provenance-aware retrieval, tool-call containment, long-horizon agent evaluation, privacy-preserving adaptation, realistic red teaming, and deployment-grade incident response.